The long-lived credential problem behind enterprise AI

AI is mainly speeding up familiar cloud risks. Datadog’s data shows high-risk vulnerabilities quadrupling and old credentials piling up. CIOs should fix identity and guardrails first.

Lack of visibility into remote end-points leaving companies vulnerable to Ransomware: Study

Datadog’s 2026 State of Cloud Security report makes an argument CIOs can take straight into a budget meeting: AI has not created a new class of cloud risk so much as it has sped up the old ones. Drawing on posture data from thousands of organisations on AWS, Azure and Google Cloud, it finds that AI-assisted development and autonomous agents are turning vulnerabilities, misconfigurations and weak identity controls into exposure at unprecedented speed and scale. Its summary is blunt: AI is changing the speed and scale of cloud security, but not its fundamentals.

The vulnerability backlog has quadrupled

The median cloud workload carried 4 high or critical vulnerabilities in February 2026. By November the figure was 16. Datadog ties the jump to a surge in reported vulnerabilities since May, driven by security-capable AI models. It cites Google Cloud’s estimate that 44.5% of cloud incidents in the first half of 2026 began with an attacker exploiting a software vulnerability. For buyers, patch capacity sized for last year’s volume will not hold. The report’s emphasis is on limiting the blast radius, so the useful question is what a compromised application can reach next.

AI services bring old credentials into new places

More than one in five organisations now use cloud-hosted AI inference services such as Amazon Bedrock, Azure Foundry Tools and Google Cloud’s Gemini Enterprise Agent Platform. Adoption rose on every cloud over the year: Google Cloud from 17% to 23% of organisations, AWS from 13% to 21%, and Azure from 11% to 16%. Convenience carries a cost. Nearly half of organisations using Gemini rely on long-lived API keys, as do 17% of those using Bedrock. Azure Foundry resources are created with long-lived keys already generated, and 97% of Azure organisations leave that behaviour unchanged. The report warns that criminals can steal such keys and resell access.

The wider credential picture is not improving fast. Active keys created in the past year exist for 43% of AWS IAM users, 41% of Google Cloud service accounts and 66% of Entra ID app registrations. The share of AWS IAM users holding keys older than three years rose from 20% to 30% between 2023 and 2026, and for Google Cloud service accounts from about 17% to 29%. Only Azure moved the other way. Single sign-on is spreading, with 63% of organisations using it exclusively for AWS console access, up from 55% in 2024, but 37% still use IAM users to some degree. The default settings in AI services are therefore a procurement question, and short-lived workload identity should be a condition of use.

The software supply chain now moves at machine speed

Sixty percent of organisations run dependency versions released less than 24 hours earlier, and those versions sit in 21% of cloud applications. A cloud workload averages 149 dependencies, more than half of them transitive, so teams cannot see much of what they run. The report describes a bot that spent 10 days probing open source projects for vulnerable GitHub Actions workflows, gained code execution in most targets and took over one repository entirely. It also describes supply-chain malware that goes after cloud credentials. Its remedy is a dependency cooldown, because the community usually identifies and contains hijacked libraries within a few hours. That is a policy a CIO can mandate without buying anything.

Guardrails exist, but enforcement is shallow

Keyless authentication for CI/CD pipelines is gaining ground unevenly. Among GitHub Actions users, 54% in AWS use it exclusively, up from 37% in 2024, against 22% in Google Cloud and 5% in Azure. In Azure, 52% mix keyless and long-lived methods. Data perimeters show the same pattern. Ninety-six percent of organisations use AWS Organisations, but only 55% implement data perimeters, up from 40% a year earlier. Most do it bucket by bucket, with 47% using S3 bucket policies, while organisation-wide controls remain rare: service control policies at 2.3% and resource control policies at 0.4%. Fourteen percent now run EC2 instances in the organisation management account, up from 9%, where one compromised workload can compromise everything.

What to weigh

This is vendor research. Datadog sells cloud security and observability, and its sample comes from its own customers. The link between AI models and the vulnerability surge is the report’s own attribution. It gives no country breakdown, so Indian CIOs cannot read their position from it. Most data was collected in September 2026, though the vulnerability trend runs to November.

What CIOs should do next?

  • Set remediation targets for high and critical vulnerabilities on the workloads that matter most, and test what a compromised application could reach.
  • Make short-lived identity the default for people, pipelines and AI services, and require a justification for any static key.
  • Start retiring keys older than three years, and block new long-lived keys through organisation-level policy.
  • Adopt a dependency cooldown policy.
  • Move data perimeters from individual buckets to organisation-wide controls, and keep workloads out of management accounts.

The report’s lesson is that the fundamentals still decide the outcome, and that attackers now exploit gaps in them far faster than before.

Share on